alibabaOfficial·6 Agent Skills Included

atrex-kernel-agent

Orchestrated GPU kernel optimization, profiling, and knowledge mining

Automates production-grade GPU kernel optimization across NVIDIA, AMD, and PPU hardware with profiling, correctness checks, and Git-isolated campaigns. Eliminates manual benchmark tuning, unreliable profiler interpretation, and lost optimization history through a structured hardware and kernel knowledge wiki. Mines past optimization sessions into validated, reusable records so each new kernel campaign starts faster and avoids known dead ends.

71|22

↗

npx skills add alibaba/atrex-kernel-agent --all -g -y

Available:

Instructs the agent on how to query the GPU knowledge wiki, attribute records correctly, and follow strict benchmark integrity and anti-caching rules during kernel optimization.

All Skills in This Repository (6)

Pure Emerald Level Indicators

📦 In Repo

alibaba

gpu-kernel-optimizer

Implement and optimize GPU kernels using PyTorch logic and DSLs.

gpu-kernel-baseline

Implement and validate baseline GPU kernels using CuteDSL or FlyDSL.

gpu-kernel-profile-optimizer

Profile GPU kernels with NCU or ROCprof and iteratively optimize bottlenecks.

gpu-kernel-partial-restart

Mask half of iteration memory and launch a subagent for GPU kernel optimization.

gpu-kernel-bottleneck-analysis

Analyzes GPU kernel performance using Roofline models and profiling data.

gpu-kernel-output-contract

Package GPU kernel implementations into a compliant generated_kernel.py module.

Official

Advanced

Frequently Asked Questions

FAQPage Schema

How to install atrex-kernel-agent?▼

Run `npx skills add alibaba/atrex-kernel-agent --all -g -y` in your terminal to install all skills in this suite globally.

What does Atrex Kernel Agent do?▼

It turns an operator specification into a measured, optimized GPU kernel by coordinating coding agents, profiling, correctness checks, and Git-isolated optimization campaigns.

Which GPUs and frameworks are supported?▼

It targets NVIDIA, AMD, and T-Head PPU hardware and supports Triton, CuteDSL, CUDA, FlyDSL, and TileLang campaigns.

What is the GPU Wiki used for?▼

It is a structured knowledge store of hardware facts and kernel optimization experience that the agent queries to pick proven techniques and avoid known dead ends.

Does it work with Claude Code and Codex?▼

Yes. It supports Claude, Codex, Qoder, and Pi as coding-agent backends and follows the standard SKILL.md format.